【Paper】 Deep Learning for Generic Object Detection, A Survey - Summary

Deep Learning for Generic Object Detection, A Survey 논문 리뷰 및 정리

Deep Learning for Generic Object Detection: A Survey

My Conclusion after I read the paper.

0. Abstract

  1. a comprehensive survey of the recent achievements
  2. More than 300 research contributions
  3. frameworks/ representation/ proposal generation/ context modeling/ training strategies/ evaluation metrics
  4. promising directions for future research

1. introduction

  • Deep Convolutional Neural Network (DCNN) called AlexNet in 2012
  • Over the past 5 years, this article attempts to track in order to gain a clearer picture
  • popular datasets/ evaluation metrics/ context modeling/ detection proposal methods.

  • many object detection previous surveys (refer to reference this paper later)
    • pedestrian detection
    • face detection
    • vehicle detection
    • traffic sign detection
    • generic object detection
  • we have limited our focus to top journal and conference papers. and limited picture detection not Video.
  • follows
    1. 20-years are summarized
    2. A brief introduction to deep learning
    3. Popular datasets and evaluation criteria
    4. the milestone object detection frameworks
    5. state-of-the- art performance
    6. future research directions

2. Generic Object Detection

2.1 The Problem

  • future challenges will move to the pixel level object detection.

2.2 Challenges

  • Accuracy Related Challenges
    • intra-class variations : color, texture, material, shape, size, poses, clothes
    • Imaging condition variations : lighting, physical location, weather, backgrounds
    • 20, 200, 91 object classes / VOC, ILSVRC, COCO is much smaller than can be recognized by humans.
  • Efficiency and Scalability Related Challenges
    • mobile/wearable devices have limited computational capabilities and storage space
    • A further challenge is that of scalability : unseen objects, unknown situations, it may become impossible to annotate them manually, forcing a reliance on weakly supervised strategies.

2.3 Progress in the Past 2 Decades

  • image-20210114173125275
  • accurate annotations are labor intensive to obtain
  • Ability to detect many object categories matches that of humans(인간 같은) 3000~3000 categories is undoubtedly an unresolved problem so far.

3. A Brief Introduction to Deep Learning

  • pass

4.1 Dataset

  • PASCAL VOC
  • ImageNet
  • MS COCO
  • Places
  • Open Images

4.2 Evaluation Criteria

  • Frames Per Second (FPS)
  • precision, and recall.
  • Average Precision (AP) -> over all object categories, the mean AP (mAP)

5. Basic Frameworks

  • image-20210114192428720
  • image-20210114201752611

© All rights reserved By Junha Song.